Deep learning operator automatic optimization system and method based on Shenwei processor

Abstract

The invention provides a deep learning operator automatic optimization system based on an Shenwei processor, and the system comprises an obtaining unit which obtains the description of a special description language definition calculation task and an optimization space; a scheduling unit which constructs an optimization space according to the description of the optimization space, schedules the calculation task description according to different optimization methods in the optimization space to generate a plurality of different calculation implementations, and outputs the calculation implementations expressed by intermediate representation; an optimization unit which is used for receiving the intermediate representation, optimizing the intermediate representation and outputting the optimized intermediate representation; an optimized space searching unit which is used for searching the optimal calculation implementation from the optimized intermediate representation; and a code generation unit which is used for translating the optimal calculation implementation into codes which can be executed on the Shenwei processor. The method can solve the problems of low optimization performance, difficult transplantation and large optimization time overhead in the prior art, is more efficient than a manual optimization technology and an automatic optimization technology, and can be conveniently transplanted to other mechanisms for use.

Description

technical field [0001] The present invention relates to an algorithm optimization system and method, in particular to a Shenwei processor-based deep learning operator automatic optimization system and method. Background technique [0002] Artificial intelligence has penetrated into all aspects of work and life. Deep learning technology has made remarkable breakthroughs in image recognition, language processing, and object detection. The Sunway supercomputing platform has also built a distributed deep learning development environment. As the depth of the deep learning model deepens and the number of parameters increases, its demand for computing resources gradually increases, and it is necessary to carry out in-depth customization and optimization of the computing operators of the model. [0003] For deep learning models, especially convolutional neural network models, the main computational resource consumption lies in computationally intensive operators, including convoluti...

AI Technical Summary

Problems solved by technology

[0006] In order to solve the above technical problems, the Shenwei processor-based deep learning operator automatic optimization system and method provided by the present invention can solve the problems of low optimization performance, difficulty in transplantation, and high optimization time cost in the prior art, and is more efficient than manual optimization. Optimization technology is more efficient than automatic optimization technology, and can be transplanted to other institutions by changing assembly primitives

Images